[RFC] Reinstate NFS exportability for JFFS2.

J. Bruce Fields bfields at fieldses.org
Sat Aug 2 17:33:38 EDT 2008


On Sat, Aug 02, 2008 at 09:42:32PM +0100, David Woodhouse wrote:
> On Sat, 2008-08-02 at 14:26 -0400, J. Bruce Fields wrote:
> > Could you add a readdirplus vfs operation which took a flag indicating
> > how much extra information you're going to need?
> 
> Actually, if we're screwing with readdir then xfs would like to know how
> much it's going to be asked to read, rather than just have the filldir
> callback return zero when it's done.

OK.

> > The three cases where readdir can be called are:
> >         - ordinary readdir, for which the existing filldir_t provides
> >           all the information needed.
> >         - nfsv3 readdirplus, where the only additional information
> >           needed is whether there's a mountpoint.
> 
> It also wants a file handle.

Oops, right.

> For which I think it just needs
> i_generation in addition to the information it already has.

Typically, right, though the filesystem's allowed some choice about what
exactly it wants to use in the filehandle.  I don't know how the various
filesystems are actually using that in practice.

> >         - nfsv4 readdir, where we may need all the stat info, acls,
> >           etc., etc.
> 
> We _might_, but most of the time we won't. It might be OK to fall back
> to the existing double-buffer hack for the cases where we _do_ need that
> extra information.

How bad is the "double-buffer hack" anyway?  Rather than have this as a
fallback case that's rarely used (hence rarely tested), it might be
simpler just to use it for everything if we're going to use it at all.

--b.

> I think a ->lookup_fh() method (which _expects_ to be called from within
> filldir, and hence does its locking automatically) is the way to go.
> 
> One alternative might just be a full lookup method with the same locking
> rules: ->lookup_locked(). That might be easier to implement, and would
> solve the problem even for the corner cases of NFSv4. Although I still
> think we'd do best to avoid populating the inode cache with _all_
> children when we do a readdirplus. 
> 
> -- 
> David Woodhouse                            Open Source Technology Centre
> David.Woodhouse at intel.com                              Intel Corporation
> 



More information about the linux-mtd mailing list